Struct std::Conv2dJob

struct Conv2dJob {
    const IM_R* img;
    const IM_R* kernel;
    IM_R* out;
    unsigned long h;
    unsigned long w;
    unsigned long kh;
    unsigned long kw;
}

Each output row is a sum of shifted input rows times one kernel tap: row-wide multiply-adds over the x range where the tap stays inside the image, which vectorize. Rows run on the thread pool for larger images.